What training an AI on yourself actually means
Training AI on yourself means uploading 8 to 15 selfies to an identity-trained photo generator. MakeAiPhotos is a selfie-trained AI photo generator that builds a personalised face model from your selfies and generates entirely new photos of you in different settings, outfits, and lighting conditions in under 30 minutes. When you upload selfies, you are giving the model enough visual data to learn your real facial features.
Training is the step where the AI learns the specific details that make you look like you: your jawline shape, how shadows fall on your cheekbones, the exact tone and texture of your skin, and the proportions between your features. Without enough photos needed to train AI on those details, the model fills them in with guesses. Guesses produce outputs that look like a relative rather than you.
Most personal AI photo tools handle the technical side automatically. Your job is to give the model enough variety to learn from, not to understand the math behind how an AI face model is built under the hood. Think of your upload batch as a small dataset: dataset size for an AI portrait model matters, but composition of that dataset matters more.
How many photos to train an AI on yourself: the 8 to 15 sweet spot
The short answer to how many photos to train AI on yourself: 8 to 15 clear selfies is the sweet spot for most current AI photo generators, including MakeAiPhotos. With fewer than 8 usable photos, the model lacks enough angle and lighting data to anchor your face reliably. Above about 20 photos, diminishing returns kick in fast, and adding near-identical photos from a burst session can hurt consistency by overweighting one expression or lighting condition.
Inside that 8 to 15 sweet spot, the ideal target for most people is 12 to 15 photos with genuine variation across at least two lighting setups and three to four angles. That dataset size for an AI portrait model is enough reference points to generalise your face accurately without flooding it with redundant data.
Why so few photos work: this is not a giant foundation model being trained from scratch. An identity-trained AI face model only has to learn one new face on top of a base model that already knows what human faces look like. 10 to 15 well-chosen photos is enough signal to teach it the specific you.
What breaks when you train on fewer than 8 photos
With fewer than 8 photos, most models start to guess at the parts of your face not well-represented in your batch. That usually shows up as eye shape drift, incorrect nose width or length, and skin texture that looks smooth in a generic way rather than natural in a way that matches your real skin.
Six nearly identical selfies taken in the same spot, same angle, same lighting are worth roughly the same as two photos to the model. Duplication does not add information to the dataset. Variety does. If your batch is small, diversify it before uploading more of the same.
Below 6 photos, expect outputs that occasionally look like a sibling or cousin rather than you. That is the model interpolating between the few faces it has seen, plus its base understanding of generic human features.
What breaks when you train on more than 20 photos
Uploading 40 or 50 photos sounds thorough. In practice, large batches usually contain many frames from the same burst session that are nearly identical, a few shots with filters or heavy beauty mode that teach the AI face model the wrong version of your face, and photos from years ago showing different hair or weight.
A large, unfiltered batch can produce outputs that look averaged, slightly plastic, or like a composite of several versions of you. Curate before you upload. Remove duplicates, filter shots, and any photo from more than two years ago where you look noticeably different.
More photos do not mean a better dataset. A clean 12-photo batch consistently beats an unfiltered 35-photo batch on likeness scores.
The photo-type breakdown that improves training quality
The key insight most guides skip: variety in photo type matters more than total count. A batch of 12 photos with three different light sources, four angles, and two distances will almost always outperform 25 photos taken from the same angle under the same lamp.
Aim for this mix across 12 to 15 photos: at least four window-light photos where your face is clearly lit without harsh shadows; two to three photos outdoors in open shade or soft daylight; two to three indoor lamp photos that differ from the window set in colour temperature; and two to three photos taken from about one metre away with your shoulders visible rather than arm's-length closeups.
The one-metre photos are the most skipped and the most important. Front-facing phone cameras use a wide-angle lens that makes your nose appear wider and your ears appear further back than they actually are. Training on only arm's-length selfies teaches the model that distorted version of your face rather than your real proportions.
Does the photo count change depending on what you want from your AI photos?
The core 8 to 15 sweet spot applies across every use case once your AI face model is trained. The same identity-trained model can produce LinkedIn headshots, dating app photos, travel content, and social media shots without retraining. What changes is how you curate the dataset, not how many photos you upload.
For LinkedIn and professional use: include at least three photos in a neutral expression, business-casual or plain clothing, and soft natural light. The model needs to learn how you look in a professional context.
For dating and lifestyle: include a range of expressions including a natural smile, and at least one photo where you are not looking directly at the camera. That variety allows the model to generate more natural-looking lifestyle scenes rather than stiff forward-facing frames.
For social media content across many styles: add an extra one to two photos in different settings, one indoor cafe-style, one outdoor. That broadens the model's understanding of how you look in varied environments.
If you want a full walkthrough of the upload, training, and generation steps, see the guide to how to make AI photos of yourself for the step-by-step flow.
Frequently Asked Questions
- How many photos do I need to train an AI on myself?
- The sweet spot for training an AI on yourself is 8 to 15 clear photos, with 12 to 15 being the practical target for most people. Fewer than 8 leaves the model without enough angle and lighting data to keep your face accurate. More than 20 usually adds duplicates that hurt consistency rather than helping. Photo variety across lighting and angles matters more than total count.
- How many photos does an AI model actually need to learn a face?
- An identity-trained AI face model is not built from scratch. It is added on top of a base model that already understands generic human faces. Because of that, 10 to 15 well-chosen photos is enough signal to teach it the specific you. The dataset size for an AI portrait model is small precisely because the base model is doing most of the work.
- What happens if I train an AI face model on too few photos of myself?
- Below about 8 clear photos, the model guesses at details it cannot see, producing eye shape drift, incorrect nose proportions, or skin that looks smooth in an uncanny way. Six nearly identical selfies count as roughly two photos in terms of training value. Diversify before adding more.
- Does it matter what types of photos I upload to train AI on yourself?
- Yes, more than the total count. A mix of window light, outdoor shade, and a few photos from one metre away rather than arm's-length only gives the model more accurate proportion data. Including two to three lighting setups and at least three to four angles is more valuable than doubling a small batch.
- Can I train an AI on old photos of myself?
- Only if they still match how you look today. Photos from two or more years ago with different hair, weight, or facial hair teach the model an outdated version of your face. If a photo would confuse someone who has not seen you in a while, remove it from the training batch.
- How often do I need to retrain the AI on new photos of myself?
- Most people retrain when their appearance changes noticeably, new haircut, beard change, significant weight change, or roughly every one to two years. Small differences in styling do not usually need a new training session if your base selfies are strong.
- Will uploading more photos always improve my AI photos?
- Not above about 20 photos. Beyond that threshold, each additional photo tends to add redundancy rather than new information. A curated set of 12 varied photos consistently outperforms an unfiltered batch of 45 taken from the same position.